Predicting gross-total resection in skull base chondrosarcoma: a multi-institutional machine learning study.

Abstract

Objective: The aim of this study was to develop and validate an anatomy-driven model that predicts gross-total resection (GTR) of skull base chondrosarcomas (SBCs).

Methods: The authors conducted a retrospective multi-institutional analysis of consecutive SBCs resected at 4 academic centers in 2000-2024. Thirteen preoperative variables of interest included tumor location (petroclival, peri-lacerum, anterior, or lateral skull base), maximal tumor diameter, internal carotid artery (ICA) encasement, compartmental extension, cranial nerve (CN) involvement, prior radiotherapy and/or surgery, synchondrosis of origin, and planned operative approach. Data were split into 75% training (n = 139) and 25% validation (n = 46) sets. Five algorithms were tuned: generalized linear model (GLM), random forest, support vector machine, k-nearest neighbors, and eXtreme Gradient Boosting. For interpretability, a multivariable logistic regression model was fitted and a nomogram for predicting GTR was built. Discrimination performance was summarized by the area under the curve (AUC), accuracy, sensitivity, specificity, and Brier score.

Results: A total of 185 consecutive SBCs were identified. Anatomical burden varied by location: cavernous sinus invasion clustered in peri-lacerum (64.5%) and petroclival (57.1%) SBCs, jugular foramen extension occurred largely in petroclival cases (38.1%), and sinonasal or orbital spread characterized anterior midline lesions. Surgical approaches mirrored this distribution, endonasal routes for peri-lacerum (80%) and petroclival (63.4%) tumors, whereas open approaches predominated for lateral tumors (84.6%). The overall GTR rate was 56%, with rates of 52.4%, 67.7%, 76.9%, and 66.7% for petroclival, peri-lacerum, lateral, and midline tumors, respectively (p = 0.186). For petroclival disease, the surgical choice was decisive (p < 0.001), with GTR rates of 72.1% for the endoscopic transpterygoid approach (ETPA), 25.7% for open surgery, and 41.7% for the midline endoscopic endonasal approach (EEA). Multivariable analysis demonstrated significantly decreased odds of GTR with graded ICA encasement (ORs 3.12 for 90°-180°, 7.41 for 181°-270°, 9.76 for 271°-359°, 8.52 for 360°, compared with no encasement), prior radiotherapy (OR 4.04), petroclival location (OR 4.72), lateral location (OR 3.29), infratemporal fossa extension (OR 2.75), and CN VII-VIII involvement (OR 2.82). The ETPA, as compared to open surgery, was independently associated with higher odds of GTR (i.e., decreased odds of subtotal resection; OR 0.22, p < 0.001), with a favorable trend for midline EEA (OR 0.37, p = 0.063). The GLM exhibited the best discrimination with an AUC of 0.838 and Brier score of 0.19 (accuracy 0.898).

Conclusions: Anatomical factors, especially petroclival tumor origin, ICA encasement, and lower CN corridors, are the principal barriers to complete resection of SBCs. The machine leaning-supported nomogram provides a reproducible preoperative tool aligning corridor choice with anatomy, improving GTR likelihood, and rationalizing adjuvant therapy.